An AI visibility audit is a repeatable check of how ChatGPT, Perplexity, Gemini and Google AI Overviews talk about your brand when buyers ask category questions. You run a fixed set of buyer-style prompts, log every answer in the same sheet, and repeat it monthly so the numbers start meaning something.
- Log four things per answer: brand appearance, position, competitors named, sources cited.
- Run every prompt 3 to 5 times in fresh sessions to avoid contradictions.
A 15-to-20-prompt baseline across ChatGPT, Perplexity and Gemini takes 2 to 3 hours and belongs before any tool subscription.
- AI often cites pages outside your Google top 10, and that gap shows which content to fix.
What an AI Visibility Audit Actually Measures
An AI visibility audit measures brand frequency, prominence, and cited authority. Every metric comes from four data points logged per answer: appearance, position, competitors named, and sources cited.
| Metric | What it tells you | How to calculate it |
|---|---|---|
| Mention rate | How often you show up at all | Runs where your brand appears ÷ total runs × 100 |
| Average position | How prominently you’re listed | Sum of your positions ÷ runs where you appear |
| Citation share | How often your own domain backs the answer | Runs citing your domain ÷ total runs × 100 |
| Share of model (share of voice) | Your slice of the conversation against competitors | Your mentions ÷ all brand mentions (you + tracked competitors) × 100 |
| Competitor share | Who owns the answers you’re missing | Each competitor’s mentions ÷ all brand mentions × 100 |
| Sentiment | How the AI describes you | Positive, neutral and negative mentions ÷ your total mentions |
A ranking report tells you where one URL sits for one keyword. These metrics track a brand across answers that pull from very different places. The clearest data so far comes from international, largely US English-language studies, not from Austria: in a comparison of 1,000 queries across five search systems, GPT-4o had a median domain overlap of 0% with Google’s top 10. For more than half the queries it cited no top-10 domain at all, while Claude and Perplexity landed at 12 to 15%.
Even Google’s own AI Overviews have drifted away from its rankings. The share of cited pages that also rank in the top 10 dropped from roughly three quarters in late 2024 to 38% in a 2026 analysis of 863,000 keywords and to 17% under a stricter method. All of these figures come from international datasets, so they show the direction of travel, not what Austrian or DACH queries look like. If leadership keeps asking about clicks, it helps to have a plan for measuring AI visibility beyond traffic.
Building Your Prompt Set Across the Buyer Journey
Aim for 25 prompts in three buckets: ten category, five comparison or use-case, and ten branded. For the very first baseline, 15 to 20 is plenty, and you grow into the full set once the monthly routine sits. Copy these templates and swap in your own terms:
- Category (10), early to mid journey: “What’s the best [category] for [audience]?”
- Comparison and use case (5), late stage: “[Your brand] vs [competitor] for [use case]?”
- Branded (10), decision stage: “What is [your brand]?” and “Is [your brand] worth it?”
Good variations for the same buckets: “Which [category] tools work for [use case]?”, “Alternatives to [competitor] for a [team size] team” and “What does [your brand] cost?”. Phrase like a buyer would type it, in full sentences with context, like “best CRM for a 10-person agency”. Sales calls and support tickets are the best wording source. And separate branded prompts in their own bucket. They return your name almost by default, so mixing them in inflates the mention rate.
Your capture sheet
Every single run gets one row in a sheet, date-stamped and shared. These columns cover the four core data points plus what the metrics above need:
| Column | What to enter |
|---|---|
| Date, platform, mode | 2026-11-03, Perplexity, Web |
| Prompt ID and run number | C-04, run 2 of 3 |
| Brand appears | Yes / No |
| Position | 1, 2, 3 or “mentioned in body text” |
| Competitors named | Comma-separated list |
| Sources cited | Domain, exact URL, citation number |
| Sentiment and accuracy | Positive / neutral / negative, plus anything factually wrong |
Running the Audit in ChatGPT, Perplexity, Gemini and AI Overviews
Run each platform as its own block, starting with Perplexity, since it puts numbered sources next to every answer and is the easiest model to audit. Use the same prompts in the same order everywhere so your rows line up across platforms.
For Perplexity, use the standard Web mode (Academic and Social filter the index differently) and note the citation number and domain for every claim about your brand. For ChatGPT, open a Temporary Chat for each run and log whether the answer shows sources at all. For Gemini, run the prompt, then hit the double-check option under the answer to see which statements Google can confirm. For Google AI Overviews and AI Mode, check both directly in Google Search, separately from the Gemini app, because they often return different brand lists.
Temporary Chat: A ChatGPT session that ignores your saved memories and doesn’t create new ones. For an audit, it’s the closest you get to asking as a stranger.
Gemini deserves a note of its own. It pulls from the live Google index plus its training data, so brands that rank well in Google tend to show up in Gemini, and weak rankers usually don’t. SEO feeds Gemini.
Neutralize the bias before you start
Memory distorts ChatGPT most. With it switched on, ChatGPT can rewrite your prompt using stored preferences such as your location before it searches, so “best agency near me” quietly becomes a personalized query. Temporary Chats avoid this, and OpenAI lets you switch memory off for good under Settings > Personalization > Memory.
Google and Perplexity don’t document location and language effects the same way, so control them yourself. Test from a clean browser profile, set the interface language to the market you’re auditing, and record your location in the sheet.
Auditing for Austria and the wider DACH market
If you sell in Austria, run from an Austrian location with German interface language for Google and Gemini, and log “AT” in a market column. Write the prompts the way Austrian buyers actually phrase them, including local terms, instead of translating your English set word for word. Selling into Germany, Switzerland or English-speaking markets too? Run separate sets per market, because each one can surface its own sources and competitors.
The published benchmarks in this article come from international datasets, so your Austrian sheet becomes the only reliable reference for that market.
How Many Repeats You Need Before Trusting the Numbers
Run each prompt at least three times per platform in separate, fresh conversations, five if you have the time, and log the majority result. One run shows what the model said once. Three to five runs show whether it says it consistently. Twenty prompts across three platforms at three runs each is 180 rows, which is why the first baseline takes an afternoon.
Cadence matters just as much. In a Swiss study from St. Gallen, only about half of the cited URLs overlapped between consecutive runs, and the authors recommend rolling observation windows of two to four weeks. A one-off audit goes stale fast, so run it monthly minimum. During a launch or when a competitor is pushing hard, add weekly checks on your five most important prompts.
Rule of thumb: Count a change as real when it holds across all repeats in one month and shows up again the next. Anything that appears in a single run or a single month is noise until proven otherwise.
Tracing Which Sources AI Actually Cites for Your Brand
Source analysis shows you which pages AI trusts for your category, and whether any of them are yours. For each prompt and model, record whether your domain was cited (yes/no), the exact URL, and its position in the source list.
Worth knowing: After a few months, a pattern usually appears. Two or three of your pages earn almost all of your citations, and they’re often not the ones you’d expect. Those pages deserve internal links and regular updates.
Third-party sites dominate. Comparison portals, analyst reports and G2-style review platforms often beat brand pages because they’re built to answer category questions directly. In an international analysis of 30 million sources, the most-cited domains across AI search fall into a few types:
- Community forums: Reddit, Quora
- Video platforms: YouTube (the single most-cited domain in AI Overviews at about 5.6% of all citations)
- Professional and social networks: LinkedIn, Facebook
- Review platforms: G2, Yelp
The platforms also differ from each other. In an academic comparison of AI search systems, AI Overviews and organic Google shared 68% of their top-100 domains, while SearchGPT overlapped only about a quarter with either and leaned on Wikipedia, Britannica, Reuters and AP News.
To find your gaps, take each category prompt and pull top 10 rankings for the matching keyword from your rank tracker. Sort the cited URLs into three buckets: you rank and get cited, you rank but don’t get cited, or a third party gets cited where you have nothing. Third bucket = to-do list, and the second usually points to structure problems.
The published overlap figures come from US and global English-language datasets, so they serve as an international comparison, and your own sheet shows what holds in Austria. For a fast first pass on the same question, here’s how to check whether AI engines cite you.
Turning Audit Findings Into an Action Plan
Each finding calls for a different fix, and the order matters: fix wrong first.
| Priority | Finding | Likely cause | What to do |
|---|---|---|---|
| 1 | Misrepresented (wrong price, old features, wrong positioning) | Outdated or conflicting info on your site or third-party profiles | Publish one clear source-of-truth page (facts, pricing, product scope), update G2, LinkedIn and directory profiles, flag the answer through the platform’s feedback button, re-test next cycle |
| 2 | Not mentioned in category prompts | You’re absent from the sources the AI trusts for your category | Build pages that answer category questions directly and get listed on the review sites and roundups you saw cited |
| 3 | Competitor dominates | The competitor owns the cited third-party sources | Target the specific URLs cited for them: comparison pages, review volume, presence in the relevant community threads |
| 4 | Mentioned but not cited | Your pages lack structural signals | Add a roughly 40-word definition block at the top, FAQPage or HowTo schema, and a named author with publish date |
Check one detail before you blame structure. When ChatGPT answers without browsing, it works from training data and shows no sources at all, so missing citations tell you nothing about your page. Missing citations signal gaps when the answer pulled from the web. For wrong claims in Gemini, the double-check feature shows which statements Google can’t confirm, which tells you where your public information is thin.
All of this works with your CMS and a few hours a month. If you’d like an outside view first, Claire’s Free GEO Audit checks your site against 70+ points that decide whether ChatGPT and other AI systems find, understand and cite it. For B2B teams, the overview of SEO and AI visibility for B2B and SaaS covers what buyers search for and the five levers that move it.
When to Move From Manual Checks to a Tracking Tool
Switch to a tool after one manual baseline completes and the monthly re-run starts costing more time than you have. The manual round tells you which prompts, competitors and markets matter, so you configure the tool properly and skip paying for tracking nobody looks at. These are the names you’ll run into most:
- Profound: enterprise-focused AI visibility for larger brands
- Peec AI: prompt tracking across platforms
- Otterly.AI: lightweight monitoring for smaller teams
- Scrunch: monitoring and optimization for AI search
Pricing changes often, and vendors count in different units, so compare on what drives the bill: prompts, platforms, refresh rate, and markets. Entry plans at the smaller tools typically cost about as much as a mid-tier SEO tool, while enterprise platforms usually quote on request. Run your shortlisted tool next to your manual sheet for one cycle, and if the numbers diverge wildly, get an explanation before you sign an annual plan.
If you’d rather hand off both the tracking and the fixes, Managed Claire (from €1,999) is a GEO and SEO service where AI agents work on making your brand visible in ChatGPT and other AI search engines.
Traffic is the other half of the picture. GA4’s native AI Assistant channel group sorts referrals from ChatGPT, Gemini, Copilot and a few others automatically, but Perplexity stays in regular referrals. Add a custom channel group or exploration that filters session source for perplexity.ai so those visits don’t vanish into the general bucket. Search Console helps less: AI Overviews and AI Mode clicks land in the regular web search report without a separate filter.
Many AI answers never produce a click, so referrals show the floor. Your audit sheet shows the rest.
Your Baseline Is the Starting Line, Not the Finish
A single audit is a snapshot. Runs two and three turn it into an early-warning system that catches a competitor gaining ground or a wrong claim spreading before it ever shows up in your pipeline. Skip the wait for the perfect tool: pick 15 prompts this week, run them through Perplexity, ChatGPT and Gemini with three repeats each, and put next month’s re-run in the calendar before you close the sheet.
FAQ
Is a manual audit reliable enough to report to leadership?
Usually, as long as you present it as directional. A baseline of 20 to 50 prompts with a few runs each shows clear patterns in mentions and citations. Tight weekly trend lines with error margins need more like 50 to 150 prompts on a fixed schedule.
What belongs in an audit report for leadership?
Five baseline numbers: mention rate, share of model against your top three competitors, citation share, the count of factual errors found, and the three source gaps you’ll fix first. Add the date and prompt count so next month compares cleanly.
How long does it take for fixes to show up in AI answers?
Give a fix two to three monthly cycles before you judge it. Updated pages need time to get crawled and picked up, and with citation sources shifting every month, one good result can still be noise.
